全文下载： 
202202017.pdf
文章编号： 1672-6987（2022）02-0111-10； DOI： 10.16351/j.1672-6987.2022.02.017
司恩鹏1， 王鲁豫2， 李浩杰1*， 杜军威1（1.青岛科技大学 信息科学技术学院，山东 青岛 266061； 2.山东省实验中学，山东 济南 250001）
摘要： 基于社区问答CQA（community-based question and answering）的知识分享已成为互联网时代的主流交互平台，然而随着大量用户参与和大量问题的涌入，普遍存在问题回复慢而领域专家又难以发现合适的问题回答的“回答饥饿”（answer hungry）现象。已有的专家推荐方法多基于提问者、问题、答案、回答者、社交网络等中的局部视角进行特征提取，并没有分析哪一类特征或特征组合是所需的推荐特征；同时，采用机器学习或深度学习进行专家推荐时，其推荐标签特征仅区别“best answer”和非“best answer”，而不能评价非“best answer”的回答者真实知识水平，存在推荐质量不高问题。本工作充分利用用户对答案的反馈评价作为答案质量细粒度评分，设计一种考虑特征组合与交互的FM回归模型进行专家推荐，并在此基础上评价每类特征在专家推荐的作用。在爬取的Stack Overflow数据集上，按真实时间序列进行专家推荐，较主流基准算法有不同程度提升，也实证了提问者、问题、答案、回答者这几类特征组合对专家推荐的价值。
关键词： 社区问答服务； stack overflow； 专家推荐； 特征融合； 因子分解机
中图分类号： TP 311.5文献标志码： A
引用格式： 司恩鹏， 王鲁豫， 李浩杰，等. 一种基于特征融合与评分反馈的CQA专家推荐方法［J］. 青岛科技大学学报(自然科学版), 2022, 43(2): 111-120.
SI Enpeng， WANG Luyu, LI Haojie, et al. A CQA expert recommendation method based on feature fusion and scoring feedback［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition), 2022， 43（2）： 111-120.
A CQA Expert Recommendation Method Based on Feature
Fusion and Scoring Feedback
SI Enpeng1， WANG Luyu2, LI Haojie1， DU Junwei1
(1.College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China；
2.Shandong Experimental High School, Jinan 250001,China)
Abstract: Community-based question and answering (CQA) knowledge sharing has become a mainstream interaction platform in the Internet era, but with the participation of a large number of users and the influx of a large number of questions, there is an answer hungry phenomenon that it is slow to answer questions and it is difficult for domain experts to find appropriate answers. Existing expert recommendation methods are mostly based on feature extraction from the local perspective of questioner, question, answer, responder, and social network, without analyzing which type of features or combination of features are the required recommendation features; meanwhile, when machine learning or deep learning is used for expert recommendation, the recommendation label features only distinguish between "best answer" and "non-best answer", but cannot evaluate the real knowledge level of "non-best answer" answerers, so there is a problem of low recommendation quality. In this paper, we make full use of users′ feedback evaluation of answers as a fine-grained score of answer quality, design a FM regression model considering the combination and interaction of features for expert recommendation, and evaluate the role of each type of feature in expert recommendation on this basis. On the crawled Stack Overflow dataset, expert recommendation by real time series has been improved to different degrees compared with the mainstream benchmark algorithm, and the value of feature combinations of questioner, question, answer, and responder on expert recommendation is also empirically demonstrated.
Key words: community-based question answering; stack overflow; expert recommendation; feature fusion; factorization machine
收稿日期： 2021-04-05
基金项目： 国家自然科学基金项目（61973180，6217072142）；山东省自然科学基金项目（ZR2019MF014，ZR2019MF033，ZR2021MF092）
作者简介： 司恩鹏（1996—），男，硕士研究生.*通信联系人.
